Short answer

Do not start with a chatbot. Start with what the customer never sees.

The order that works: sort incoming enquiries, summarise them, and put a finished draft in front of the person who will reply. No automation facing the customer in step one.

The one rule that decides whether it goes well: a human reads it before anything leaves the building. That removes most of the time without shifting responsibility.

Why the chatbot is rarely the answer

Everyone has met one. You type your question, get an answer about something else, rephrase it, get the same answer, and then search for the word “person” until something happens. By that point the enquiry is harder to solve than if no bot had existed at all.

This is not an argument against the technology. It is an argument against where you put it. A chatbot is the most visible point in the whole customer journey, which makes it the place where every mistake costs the most.

A worked example, not a measurement: if you get thirty enquiries a week, a bot might save a few hours. A single enquiry handled badly in a small business often costs more than that, because customers usually know you personally and expect someone to reply. That asymmetry is the whole reason I almost always advise against starting there.

In a company with thousands of enquiries a month the sum looks different. But those companies rarely ask me.

Three steps, in the right order

Step one: sort. Incoming enquiries are tagged and passed to the right person, with a three-line summary at the top. Nothing goes out to the customer. This sounds small and is often the single biggest time saving in the whole chain — especially where several people read the same inbox and everyone starts again from the beginning.

Step two: draft. Once an enquiry is sorted, a suggested reply is ready. The person answering reads it, changes what needs changing and sends it. The time saving here is large, because the hard part of a customer reply is rarely writing it. It is getting started.

Drafts only become good if they are built on what you actually reply. That work — collecting earlier replies, the FAQ and what sits in the head of the person who has answered for five years — is what takes the time in the assignment. It is also useful in itself: many people discover during that work that they answer the same question differently depending on who happens to read the email.

Step three, sometimes: a direct reply to the customer. And then only in the narrow category of questions where the answer is unambiguous, harmless and the same every time — opening hours, delivery status, where to find an invoice. Never on anything involving money, contracts, complaints or anyone’s personal details.

The rule that decides everything

A human reads it before anything leaves the building.

It is the only rule I hold to in virtually every customer-service job for smaller companies, and it is worth understanding why — because everyone wants to negotiate it once the time saving starts to look tempting.

The reason is that the technology is sometimes wrong, and it is wrong with great confidence in its tone. It will write a well-phrased, assured and completely incorrect answer about your returns policy, and nothing in the text shows that it is wrong. The only person who can see it is someone who knows the business.

The objection is usually that the approval step eats up the saving. It doesn’t. Reading and sending a finished draft is considerably quicker than writing the reply from scratch. You keep most of the saving and avoid the whole risk.

When a bot is actually right

There are situations where it is, and they share three things.

The volume is high enough that people cannot keep up. Not “we would like to save time”, but “we don’t reply to everything today”. Then the alternative to a bot is not a person — it is an unanswered email, and the comparison is entirely different.

The questions are genuinely repetitive. If eighty per cent of enquiries are the same five questions, there is something to automate. If every enquiry is unique, there is not, however much you wish otherwise.

There is a way out. A bot without a clear, quick and visible route to a person is a barrier, not a service. That route should be in the first reply, not after five attempts.

Seasonal businesses are the clearest case of all: when the workload varies by a factor of ten over the year, you cannot staff for the peak, and then automated first contact is often the only realistic option.

When to hold off entirely

There are customer services where AI should not come in, or not yet.

The enquiries are few. If you only get a few enquiries a week there is little time to win, and the groundwork can take longer than you save.

Every enquiry is different. Then there is no pattern to build the drafts on.

The enquiries are sensitive. Complaints, claims and anything that calls for empathy should go to a person from the start.

You do not know where the personal data ends up. Find that out first, see AI and GDPR.

And if the answers are not written down anywhere, the first step is to write them down, not to buy something.

What it costs in time

It is two different things: the time to set it up, and the time it takes every week afterwards.

Setting it up. Most of it is rarely the technology. It is gathering what you actually usually reply — earlier emails, an FAQ, what is in the head of the person who has been answering for a long time — and choosing one type of enquiry to start with. How long it takes depends on how much is already written down, and I give no figure here, because it would be a guess.

Every week. Someone still reads every draft, and that is the point. The saving is therefore part of the time, not all of it. In the first months some time also goes on following up how much the staff change in the drafts, more on that below.

If you want help setting it up there is the service AI customer service for business.

What to measure

Not how many enquiries were “handled automatically”. That number can be pushed as high as you like by lowering quality, and people do it without noticing.

Measure three things instead. Time to first reply — it should go down. The share of enquiries that come back because the reply did not do its job — it should not go up. And how much the staff change in the drafts, which is the most useful measure of all: if they change a lot, the material is poor; if they change nothing at all, check that they are actually reading it.

The third number is also the one that shows when you are ready to let something go straight to customers — if that category of questions is hardly edited over three months.

The most important thing

If you only take one thing away: the part of customer service worth automating first is the part the customer never sees.

Sorting and drafts give most of the time saving and almost none of the risk. The chatbot does the opposite — a lot of visibility, a moderate saving, and every mistake in front of the customer.

What such a build costs and how it works is set out on the page about AI automation. And before anything is automated, find out where the personal data ends up — that part is covered here.

Best wishes, Mikael GlännströmWellbeing coach, author and father of two